I think IMO results were driven by general purpose advances, but I agree I can’t conclusively prove it because we don’t know details. Hopefully we will learn more as time goes by.
An informal argument: I think currently agentic software engineering is blocked on context rot, among other things. I expect IMO systems to have improved on this, since IMO time control is 1.5 hours per problem.
I think non-formal IMO gold was unexpected and we heard explicitly that it won’t be in GPT-5. So I would wait to see how it would pan out. It may not matter in 2025 but I think it can in 2026.
I think it is important to note that Gemini 2.5 Pro Capable of Winning Gold at IMO 2025, with good enough scaffolding and prompt engineering.
Do you have any Solomonoff inductor you know? I don’t, and I would like an introduction.
Ethan Mollick’s Using AI Right Now: A Quick Guide from 2025-06 is in the same genre and pretty much says the same thing, but the presentation is a bit different and it may suit you better, so check it out. Naturally it doesn’t discuss Grok 4, but it also does discuss some things missing here.
Anthropic does have a data program, although it is only for Claude Code, and it is opt in. See About the Development Partner Program. It gives you 30% discount in exchange.
CloudMatrix was not, but Huawei Ascend has been there for a long time, and was used to train LLM even back in 2022. I didn’t realize AI 2027 predated CloudMatrix but I still think ignoring China for Compute Production was unjustified.
This is a good argument and I think it is mostly true, but this absolutely should be in AI 2027 Compute Forecast page. Simply not saying a word about the topic makes it looks unserious and incompetent. In fact, that reaction happened repeatedly in my discussion with my friends in South Korea.
I know cyber eval results are underelicitation. Sonnet 4 can find zero day vulnerabilities, we are now in process of disclosing. If you can’t get it to do that it’s your skill issue.
Preordered ebook version on Amazon. I am also interested in doing Korean translation.
I disagree on DeekSeek and innovation. Yes R1 is obviously a reaction to o1, but its MoE model is pretty innovative, and it is Llama 4 that obviously copied DeepSeek. But yes I agree innovation is unpopular in China. But from interviews of DeepSeek founder Liang Wenfeng, we know DeepSeek was explicitly an attempt to overcome China’s unwillingness to innovate.
Maybe we are talking about different problems, but we found instructing models to give up (literally “give up”, I just checked the source) under certain conditions to be effective.
Our experience so far is while reasoning models don’t improve performance directly (3.7 is better than 3.6, but 3.7 extended thinking is NOT better than 3.7), they do so indirectly because thinking trace helps us debug prompts and tool output when models misunderstand them. This was not the result we expected but it is the case.
I happen to work on the exact sample problem (application security pentesting) and I confirm I observe the same. Sonnet 3.5/3.6/3.7 were big releases, others didn’t help, etc. As for OpenAI o-series models, we are debating whether it is model capability problem or model elicitation problem, because from interactive usage it seems clear it needs different prompting and we haven’t yet seriously optimized prompting for o-series. Evaluation is scarce, but we built something along the line of CWE-Bench-Java discussed in this paper, this was a major effort and we are reasonably sure we can evaluate. As for grounding, fighting false positives, and avoiding models to report “potential” problems to sound good, we found grounding on code coverage to be effective. Run JaCoCo, tell models PoC || GTFO, where PoC is structured as vulnerability description with source code file and line and triggering input. Write the oracle verifier of this PoC: at the very least you can confirm execution reaches the line in a way models can’t ever fake.
OpenAI wasted a whole year between GPT-3 and GPT-4. (Source: Greg Brockman said this in an OpenAI developer event.) So yes, I think OpenAI was 12+ months ahead at one time.
I think if you weren’t carefully reading OpenAI’s documentation it was pretty easy to believe that text-davinci-002 was InstructGPT (and hence trained with RLHF).
Not only was it easy, in fact many people did (including myself). In fact, can you point a single case of people NOT making this reading mistake? As in, after January 2022 instruction following announcement, but before October 2022 model index for researchers. Jan Leike’s tweet you linked to postdates October 2022 and does not count. The allegation is that OpenAI lied (or at the very least was extremely misleading) for ten months of 2022. I am more ambivalent about post October 2022.
This comment is probably not very useful, but my first thought was: “we invented a polygraph for AI!”.
2025-08 update. Anthropic now defaults to (you can opt out) using your chats for AI training, see for example https://techcrunch.com/2025/08/28/anthropic-users-face-a-new-choice-opt-out-or-share-your-data-for-ai-training/